Gemini local citations favor business websites over directories

If you run local SEO as if directories still do most of the heavy lifting, this study says you are late. Across 14,472 citations pulled from 1,487 local recommendation queries in 50 major U.S. metros, Gemini pointed to companies' own websites nearly 60% of the time. That is the headline. The bigger story is the instability underneath it. The same query repeated twice shared only about 40% of cited sources, and Gemini surfaced the same top business only about 7% of the time. For local brands, that changes the job. Winning AI search is less about chasing one ranking and more about making your site the clearest source of truth, then measuring whether that advantage survives across prompts, metros, and models.

What did the study actually prove about Gemini local citations?

It proved that business websites matter more than many local teams assumed. In this dataset, 59.9% of Gemini's citations pointed to a business's own website, which was more than directories, review platforms, and forums combined. If your local strategy still treats the website as a brochure and third-party listings as the real growth engine, that model looks outdated.

The scale matters here. The researchers analyzed 14,472 citations across 1,487 queries spanning 50 large U.S. metro areas and 10 local service categories. These were recommendation-style searches such as "best plumber near me" or equivalent local-intent phrasing, so the findings are directly relevant to high-intent local discovery.

There is an important nuance. Citation share is not the same thing as a proven ranking factor. Gemini may choose a business for multiple reasons, then cite the business's site to verify details such as services, hours, pricing, or location. Even with that limitation, the pattern is hard to ignore: when Gemini grounds a local answer, it usually wants to land on the company's own pages.

That lines up with what we have already argued in local AI search now treats your website as the source of truth. In practice, AI systems still need a clean, crawlable, specific version of your business. If your site is vague, thin, or outdated, the model has less reliable material to work with.

Why is the website win only half the story?

Because the website is winning citations, not replacing every other local signal. The study does not say directories, reviews, or community discussion stopped mattering. It says Gemini often ends up citing the business site as the page it can quote. That is a different claim, and it matters.

One example from the same dataset makes this clear. Recommended businesses averaged 4.75 stars, and 97% of them were rated 4.0 or higher. But a plain Google local baseline actually averaged even higher, at 4.84 stars. In other words, AI is not inventing a brand-new quality threshold. It is mostly inheriting a quality bar that already exists in local search.

That is why local AI visibility is still a systems problem. Your website may be the citation destination, but your reputation, category relevance, and supporting signals still influence whether you are easy to trust. A weak profile with a strong site can limit you. A strong profile with a weak site can do the same.

For a plumber, this might mean a service area page that clearly states neighborhoods, emergency availability, and real service types. For a dentist, it may mean detailed treatment pages that align with healthcare directories the model already recognizes. For a personal injury lawyer, it could mean practice-area pages that match legal-specific sources such as Best Law Firms, Super Lawyers, or Justia. The pattern changes by vertical, but the lesson stays consistent: your site has to make the business legible.

That is also where GEO technical audits become practical, not theoretical. If AI systems are repeatedly citing business pages, then crawlability, page structure, consistency, and technical discoverability stop being back-office details. They become visibility inputs.

Why is grounding drift the bigger issue for GEO teams?

Because instability changes how you measure success. Grounding drift is the tendency for an AI engine to pull a different source set for the same question across repeated runs. In practice, that means the answer you see right now may be real, but it may not be repeatable five minutes later.

In the study, different phrasings of the same underlying question shared cited sources only about 40% of the time on average. More strikingly, identical repeated calls with zero wording change still produced only about 40% to 46% domain overlap. The same top business appeared only about 7% of the time in Gemini's repeated-query tests.

That is not what classic local search looks like. In the control test, Google's local pack returned the same top listing about 90% of the time. So the volatility is not simply "search is noisy." It is a generative-answer problem.

For marketers, this breaks a lot of inherited habits. A screenshot of one good answer is not evidence that you are winning. A screenshot of one bad answer is not evidence that you disappeared. Both may be true snapshots, and both may be misleading if you treat them as stable rankings.

A concrete example helps. Imagine a multi-location HVAC brand checking "best AC repair in Phoenix" on Monday morning and seeing its own site cited. A second run could pull a different source mix, mention a different competitor, or cite Reddit plus a directory instead. If the team reports success or failure from one manual check, it is building strategy on variance.

That is exactly why AI Visibility matters in GEO. You need repeated prompts, cross-model comparisons, and trendlines over time, not isolated anecdotes. We made a similar point in AI search visibility depends on new signals: measurement has to move beyond rankings and into citations, recommendation rate, and answer framing.

What changed by category, metro, and model?

The study shows there is no universal local AI playbook. Citation behavior changed materially by vertical, and the source ecosystem behind a lawyer query did not look like the one behind a dentist or auto repair query.

Personal injury lawyer searches leaned toward business sites plus legal-specific directories, with almost no meaningful social presence. Dentist queries leaned on healthcare-specific marketplaces such as Zocdoc, Healthgrades, and Delta Dental. Auto repair was the most Reddit-dependent vertical measured. Home trades such as plumbing, HVAC, electrical, and pest control spread more evenly across own sites, general directories, and some social sources.

That matters because a generic checklist can waste time. If you tell every local business to "get on the big directories" and call it a day, you will miss the category-specific sources Gemini actually uses. A dentist ignoring healthcare marketplaces is making a different mistake than a locksmith ignoring site quality, and an auto repair shop may be fighting a stronger community-discussion dynamic than either of them.

Metro variation matters too. Some national platforms appeared across all 50 metros, which makes them baseline visibility infrastructure rather than a differentiator. A second layer of sources changed by region. The study also found that only 2% of the 4,410 unique businesses Gemini named appeared in more than one metro. That means the long tail is still alive. Local AI search is not only a story about giant aggregators.

Then there is the model gap. On the same 1,487 queries, Gemini and ChatGPT shared cited domains only about 8% of the time and named the same top business only 4.2% of the time. In the original comparison, Gemini heavily favored business websites, while ChatGPT relied much more on forums and directories. That is a sharp reminder that "AI visibility" is not one surface. It is multiple engines with different sourcing behavior.

If you want to inspect which pages and domains keep shaping those answers, Source Analysis is the right layer to look at. It helps teams move from "Were we mentioned?" to "What source ecosystem is causing that answer?" That is the difference between observation and diagnosis.

BotRank's Take

The most useful lesson here is not "websites are back." They never left. The useful lesson is that local AI search rewards businesses that are easy to verify, then punishes teams that measure visibility too casually. When the same query can produce a different source mix on the next run, manual spot checks become a false comfort. You do not need more screenshots. You need a measurement system.

That is where BotRank's AI Visibility feature fits naturally. It lets teams create reusable prompts, run them across multiple LLMs, compare whether the brand is recommended, and track changes over time. In this context, the value is not vanity reporting. It is operational clarity. You can see whether your site is actually becoming the cited source, whether competitors dominate certain metros or prompt types, and whether a promising result holds up across repetition. That is the kind of evidence local GEO needs.

How should local brands respond now?

They should treat local AI search as a retrieval and trust problem, not as a mystical new channel. The study points to a practical response: strengthen the business site, support it with the right off-site signals, and measure performance as a pattern instead of a one-off answer.

• Upgrade service and location pages. If Gemini often cites business websites, then generic copy is a direct liability. Spell out services, locations, constraints, proof points, and contact details clearly.
• Make technical accessibility boring and reliable. Crawlable pages, clean internal linking, and consistent business information matter more when the site is the page AI is likely to quote. This is also why local AI search is choosing businesses before customers click is such an important shift to understand.
• Map your real citation ecosystem by vertical. A lawyer, dentist, and auto repair shop do not need the same support sources. Look for the directories, marketplaces, and community surfaces your category actually triggers.
• Track variability across prompts and models. Do not judge performance from one phrasing. Measure recommendation rate, source overlap, and competitor presence over time.
• Turn findings into actions. Once you know which pages are weak or which signals are missing, push that work into a repeatable workflow with recommendations and execution planning.

There is also a limit worth stating clearly. This study covers 50 U.S. metros, 10 service categories, and recommendation-style local queries. It is strong evidence, but it is not a universal law for every country, every intent, or every AI product state. Teams should use it as a directional map, then validate against their own prompts and markets.

If that sounds familiar, it should. The broader theme is the same one we explored in how brands build trust across the new search journey: users now assemble confidence across multiple surfaces before they act. AI answers are one layer, not the whole path. Your brand has to stay consistent across them.

FAQ

Does this mean directories no longer matter?

No. Directories still matter as validation and category signals, and some verticals depend on them heavily. The takeaway is that they are not a substitute for a strong business website.

Is Google Business Profile enough for local AI visibility?

No. A strong profile helps, but this study suggests Gemini often cites the website itself when grounding local answers. If the site is weak, incomplete, or inconsistent, that can still limit visibility.

Why does Reddit still show up so much in local AI search?

Because community discussion can function as trust evidence, especially in recommendation-style queries. In this dataset, Reddit alone earned 13.7% of Gemini's citations, which beat the combined local-service-directory category.

How often should teams check local AI visibility?

More than once, and in a structured way. Because repeated queries can produce meaningfully different answers, the right cadence is ongoing tracking across prompt sets, metros, and models rather than occasional manual checks.

The simple takeaway is this: if you want to earn local AI visibility, make your own site the clearest answer candidate and stop measuring success like a static ranking report. If you want to see whether that work is actually changing how models cite and recommend your brand, BotRank gives you a way to track it with evidence instead of guesswork.